改进RetinaNet的刺梨果实图像识别

为实现加工车间刺梨果实的快速识别,提出一种基于改进的RetinaNet刺梨果实图像的识别方法。基于RetinaNet的模型,对RetinaNet框架中Focal loss的bias进行改进,使其能根据不同的情况控制bias的取值,再运用维度聚类算法找出Anchor的较好尺寸并匹配到相对应的特征层,对卷积神经网络结构进行优化。通过改进RetinaNet目标检测算法对7426幅刺梨果实图像进行检测识别,并与原始RetinaNet目标检测算法对比。试验结果表明:改进的RetinaNet网络模型识别方法对6类刺梨果实的识别率分别为99.47%、91.42%、96.92%、90.92%、96.89%和93.53%,平均识别率为94.86%;相对于原始RetinaNet目标检测算法,改进算法的识别准确率提高4.21%,单个刺梨果实检测时间由60.99 ms缩减到57.91 ms,检测时间缩短5.05%。本文改进算法对加工车间刺梨果实的识别具有较高的正确率和实用性。

Image recognition of Rosa roxburghii fruit by improved RetinaNet

In order to realize the quick recognition of Rosa roxburghii Tratt fruit in the environment of processing workshop,it was proposed a recognition method based on the improved RetinaNet Rosa fruit image.Based on the model of RetinaNet,the bias of loss had been improved so that it could control the value of bias according to different situations,and then it was used the dimension clustering algorithm to find out the better size of anchor and match it to the corresponding feature layer to optimize the structure of convolution neural network.Through the identification of 7426 Rosa roxburghii Tratt fruit images,the experimental results showed that the improved network model identification method could identify 6 types of Rosa roxburghii Tratt fruits respectively 99.47%,91.42%,96.92%,90.92%,96.89%and 93.53%respectively,with an average recognition rate of 94.86%.The recognition accuracy of the improved algorithm was increased by 4.21%.The detection time of single fruit was reduced from 60.99 ms to 57.91 ms,which reduced 3.08 ms,and the detection time was shortened by 5.05%.The improved algorithm has higher accuracy and practicability for Rosa roxburghii fruit recognition in processing workshop.